Docs
Security model
Connectors are untrusted code. Here is how the host keeps them to what their manifest declares.
Connectors are untrusted code. The host is built so that a connector can only do what its manifest declares and the user has agreed to.
Sandbox
Each call runs in a fresh wasmtime store, so no state is left over between calls. A component gets no preopened directories, environment variables, arguments, stdio or sockets: only clocks, randomness, logging and the host's HTTP. The host caps wall-clock time (30 s) and memory (128 MiB); a manifest's limits can only lower these.
Network allowlist
Outbound HTTP goes only to the origins in permissions.http.allow, which must be https unless an entry opts out. The host resolves each name once, refuses private, loopback and link-local addresses, and connects to the checked address, so DNS rebinding cannot reach your network. Redirects are not followed for the guest.
Credentials never reach the guest
API keys and OAuth tokens live in the age-encrypted secrets.age. The host adds them to a request after the guest has built it, and refreshes OAuth tokens itself. Any authorization, cookie or proxy-* header the guest sets is dropped, and logs never record headers or query strings.
Secrets are bound to the connector's name, auth schemes and HTTP allowlist. A different component installed under the same name that would send them elsewhere sees none and must be signed in again.
Remote MCP servers
A remote connector is a server you chose by URL, so its origin is trusted and it sees every call made through it. Its credential goes to its origin only and is bound to its URL.
Built-ins and the management API
Only the copies of manage, cron and workflows embedded in the binary may reach the management API, and only as the profile of the call. Nothing can be installed under their names.
Strict manifests
A component is installed only if its manifest validates against the schema, every tool, resource and prompt validates against the official MCP schemas of both revisions, and every tool schema compiles with no remote $ref.
Inbound access
MCP clients authenticate with the owner's bearer token, stored only as a hash and compared in constant time, or with an OAuth 2.1 access token for one profile. Clients register without a secret and must use PKCE (S256). A revoked grant stops working at once, and requests from an Origin that is not allowed are refused.
All downstream credentials belong to one owner, the person who runs the server. Every MCP client that holds the token acts as that owner.
Last updated November 21, 2026 · This page as Markdown · Full reference on GitHub